feat: add LiteLLM judge api_type

[prodmanpd]

What does this PR do?

Adds a litellm judge api_type, so the ClawBench judge can reach any of 100+ providers (OpenAI, Anthropic, Gemini, Bedrock, Vertex, Azure, Groq, ...) through a single call instead of a per-wire-format raw-HTTP function. litellm.completion natively speaks the OpenAI / Anthropic / Gemini / Bedrock protocols, so it collapses the existing per-api_type handlers into one path, and drop_params=True keeps a single call portable across providers.

Changes:

• src/clawbench/runner/judge.py and judge_llm.py - new _call_litellm(...) + a litellm branch in each judge's dispatch (mirrors the sibling _call_openai_chat / _call_anthropic_messages).
• src/clawbench/runner/run_support/api_preflight.py - litellm preflight validation (a tiny call before containers run).
• src/clawbench/tui.py - litellm added to API_TYPES + a LiteLLM provider preset.
• pyproject.toml - optional extra litellm = ["litellm>=1.85.0,<2.0"] (lazy-imported, so the base install is unaffected).
• tests/test_judge_litellm.py - dispatch / kwargs / preflight / import-error coverage.

Additive and opt-in: existing api_type handlers are untouched; litellm is an optional dependency imported lazily inside _call_litellm, with a clear ImportError pointing to pip install clawbench[litellm] if the api_type is used without it.

Corpus

• v2
• v1
• both
• not applicable

Test plan

• Verified against the repo's frozen uv environment:
  • uv run --frozen pytest tests/test_judge_litellm.py -> 6 passed (dispatch in both judges, drop_params=True, credentials forwarded only when set, preflight, and the ImportError path).
  • uv run --frozen ruff check . -> All checks passed!; uv run --frozen ruff format --check . -> clean; uv run --frozen pyright <changed files> -> 0 errors.
  • Live E2E: ran the real judge_llm.judge_request with api_type: "litellm" routed through a LiteLLM proxy (gpt-4o-mini). Aligned instruction -> {"match": true, ...}; contradicting instruction -> {"match": false, "reason": "Instruction specifies 'BLUE shirt', request body shows 'item: red shirt'"}. This exercises the full chain: judge_request -> api_type dispatch -> _call_litellm -> litellm.completion -> provider -> parsed verdict.
  • Pin resolves: litellm>=1.85.0,<2.0 installs against the current latest stable (1.97.0).

[Perry2004]

Thanks for the contribution.

Could you clarify the use case for adding a LiteLLM API type? ClawBench already have support for OpenAI, Anthropic, and Google GenAI API types and custom URLs, which should cover almost all cases calling LLM APIs I believe.

Can you explain the motivation why a LiteLLM is needed for calling LLM APIs for the judge.